Orchard vs OpenAI Whisper API.

OpenAI charges $0.006 per minute flat on the Whisper API, no tier discount and no included quota. Orchard sits at $0.00042 per minute on the Pro plan — same minute, same Whisper-class accuracy, 14× less. For a single team transcribing 30,000 minutes a month (a typical podcast post-prod or call-analytics workload):

Where Orchard is meaningfully cheaper than Whisper-direct: anything past a few hundred minutes a month. The crossover point where Orchard beats OpenAI even on the$1/yr Hobby plan is roughly ~17 minutes a month. Past that, subscription saves you money vs Whisper pay-as-you-go.

Orchard runs on whisper.cpp with Core ML acceleration, served from our own cluster. The default model is orchard-stt-v1 — a tuned Whisper-large derivative. On the same English/Spanish test sets we use internally, WER lands within a percentage point of the OpenAI hosted endpoint, with a tilt in our favour on rioplatense Spanish where we ship a fine-tuned variant.

• English (LibriSpeech clean): parity ±1% WER.
• Spanish (CommonVoice ES): parity ±1.5% WER.
• Spanish (rioplatense, internal): +4% absolute over hosted Whisper (we fine-tuned on porteño speech).
• Multispeaker podcasts: with diarization on, Orchard returns speaker labels Whisper doesn't have at all.

OpenAI's Whisper API is single-stream per request. Orchard shards long audio across the cluster, so a 60-minute podcast lands in under 90 seconds wall time — that's ~40× real-time end-to-end, including upload and post-processing. The sync endpoint targets short utterances (under 60 s, voice agent territory), and the async endpoint with webhooks handles everything else.

• Sync POST /v1/audio/transcriptions: ~150 ms p50 on a 5 s clip.
• Async POST /v1/transcriptions/upload: 60 min audio → ~90 s wall.
• Webhooks: we POST the result to your URL when ready (Whisper has no async).

Whisper API is a single endpoint: audio in, text out. Anything else — diarization, synthesis, voice cloning — is a separate provider, separate billing, separate code path. Orchard ships three products on the same balance:

Because we mirror the OpenAI Whisper request/response shape, the official SDK works against Orchard with a base-URL swap. No fork, no shim, no waiting for an SDK update:

import OpenAI from "openai";

const client = new OpenAI({
- baseURL: "https://api.openai.com/v1",
- apiKey:  process.env.OPENAI_API_KEY,
+ baseURL: "https://api.orchardrun.com/v1",
+ apiKey:  process.env.ORCHARD_API_KEY,
});

const transcription = await client.audio.transcriptions.create({
  file:  fs.createReadStream("podcast.mp3"),
  model: "whisper-1",   // alias accepted, routed to orchard-stt-v1
});

Same call, same response shape ({ text, language, duration, segments[] }), same SDK ergonomics. If you were already using OpenAI for STT, the migration is a single env var change in your CI/CD pipeline. Median migration time across customers who've done it: under an hour.

If you're already deep in the OpenAI ecosystem (Assistants API, GPT-4 vision, the rest of the stack billed on one key) and your STT volume is genuinely tiny — under ~50 minutes a month — staying on Whisper saves you the cognitive load of a second vendor. The $0.006/min Whisper rate works out to under a dollar a month at that volume, and the lack of diarization or TTS may not matter to your use case.

Past that point — high-volume workloads, anything multilingual or Spanish-heavy, anything that needs diarization or TTS bundled, or anything that benefits from async + webhooks — Orchard is the economically obvious choice.